Search results

mx.sum over a non-last axis returns all zeros for certain large tensors (e.g. [1, 32, 4096, 131072])

bug

Status: Open.

#3784 In ml-explore/mlx;

· aidiffuser opened on Jun 30, 2026

[BUG] JACCL segfaults in ibv_reg_mr (null PD) when RDMA absent; distributed_config says "RDMA not enabled" despite rdma_ctl status=enabled, survives cold boot (macOS 26.5.1 / 25F80)

bug

distributed

Status: Open.

#3777 In ml-explore/mlx;

· andrewkriley opened on Jun 29, 2026

[Feature]: Supporting normal python datatypes in the grad method obtained from mlx.core.value_and_grad

enhancement

low priority

Status: Open.

#3774 In ml-explore/mlx;

· aaishwarymishra opened on Jun 26, 2026

Feature Request: Least Squares Solver

enhancement

low priority

Status: Open.

#3773 In ml-explore/mlx;

· danlee2002 opened on Jun 26, 2026

Metal fp_qmv_impl: out_vec_size < 8 branch uses raw scale byte instead of dequantize_scale → wrong mxfp4 matvec for output dim < 8

bug

Status: Open.

#3762 In ml-explore/mlx;

· jax-0n-git opened on Jun 24, 2026

[BUG] ring and jaccl both fail to connect (error 60/65) on 4-node M3 Ultra cluster — IP network verified routable, full mesh confirmed

distributed

Status: Open.

#3755 In ml-explore/mlx;

· rogerstom-lgtm opened on Jun 23, 2026

mx.compile: assigning an elementwise expression to a negative-strided slice writes only one element

bug

Status: Open.

#3716 In ml-explore/mlx;

· cetagostini opened on Jun 18, 2026

A19 (iPhone 17 Pro) GPU returns numerically wrong results in float32 workloads — deterministic HF corruption in a Demucs separation pipeline; M2/M3 clean

Status: Open.

#3702 In ml-explore/mlx;

· IARFLOW opened on Jun 16, 2026

[BUG] MLX doesn't publish iOS-compatible wheels

enhancement

python

Status: Open.

#3665 In ml-explore/mlx;

· freakboy3742 opened on Jun 12, 2026

head_dim=256 unfused SDPA makes 100K+ context prefill impractical — real-world evidence for reviving #3293

Status: Open.

#3658 In ml-explore/mlx;

· hojin12312 opened on Jun 11, 2026

[Performance] conv_general 5D (3D convolution) is 2-5x slower than equivalent per-frame 2D conv

performance

Status: Open.

#3625 In ml-explore/mlx;

· kimjj81 opened on Jun 4, 2026

mx.fast.metal_kernel: Add support for compiler options (-fmetal-math-mode, integer template parameters, Metal 4 Tensor types

enhancement

Status: Open.

#3592 In ml-explore/mlx;

· rajveer43 opened on May 27, 2026